Skip to content

Commit 0296e02

Browse files
welcome page updates (#4)
1 parent c0aca33 commit 0296e02

File tree

1 file changed

+13
-4
lines changed

1 file changed

+13
-4
lines changed

sae-viewer/src/welcome.tsx

Lines changed: 13 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -19,9 +19,9 @@ export default function Welcome() {
1919
{feature: {atom: 25 * GPT4_ATOMS_PER_SHARD + 8, autoencoder: AUTOENCODER_FAMILIES['v5_latelayer_postmlp'].get_ae({})},
2020
label: "police reports, especially child safety", description: "safety incidents especially related to children"},
2121
{feature: {atom: 9 * GPT4_ATOMS_PER_SHARD + 44, autoencoder: AUTOENCODER_FAMILIES['v5_latelayer_postmlp'].get_ae({})},
22-
label: "price increases", description: "ends of phrases describing commodity/equity price increases"},
22+
label: "price changes", description: "ends of phrases describing commodity/equity price changes"},
2323
{feature: {atom: 17 * GPT4_ATOMS_PER_SHARD + 33, autoencoder: AUTOENCODER_FAMILIES['v5_latelayer_postmlp'].get_ae({})},
24-
label: "ratification (multilingual)", description: "ends of phrases describing commodity/equity price increases"},
24+
label: "ratification (multilingual)", description: "ratification (multilingual)"},
2525
{feature: {atom: 3 * GPT4_ATOMS_PER_SHARD + 421, autoencoder: AUTOENCODER_FAMILIES['v5_latelayer_postmlp'].get_ae({})},
2626
label: "would [...]", description: "conditionals (things that would be true)"},
2727
{feature: {atom: 63 * GPT4_ATOMS_PER_SHARD + 8, autoencoder: AUTOENCODER_FAMILIES['v5_latelayer_postmlp'].get_ae({})},
@@ -56,6 +56,9 @@ export default function Welcome() {
5656
{feature: {atom: 733, autoencoder: AUTOENCODER_FAMILIES['v5_l8_postmlp'].get_ae({
5757
num_features: '32768', num_active_features: '8'
5858
})}, label: "Patrick/Patty surname predictor", description: "Predicts surnames after Patrick"},
59+
{feature: {atom: 64464, autoencoder: AUTOENCODER_FAMILIES['v5_l8_postmlp'].get_ae({
60+
num_features: '131072', num_active_features: '32'
61+
})}, label: "things that are unknown", description: "things that are unknown"},
5962
{feature: {atom: 56907, autoencoder: AUTOENCODER_FAMILIES['v5_l8_postmlp'].get_ae({ // similar to 33248
6063
num_features: '131072', num_active_features: '32'
6164
})}, label: "words in quotes", description: "predicts words in quotes"},
@@ -80,7 +83,13 @@ export default function Welcome() {
8083
{heading: 'Safety relevant features (found via attribution methods)', heading_type: 'h3', feature: null, label: ''},
8184
{feature: {atom: 64840, autoencoder: AUTOENCODER_FAMILIES['v5_l8_postmlp'].get_ae({
8285
num_features: '131072', num_active_features: '32'
83-
})}, label: "profanity", description: "activates in order to output profanity"},
86+
})}, label: "profanity (1)", description: "activates in order to output profanity"},
87+
{feature: {atom: 104813, autoencoder: AUTOENCODER_FAMILIES['v5_l8_postmlp'].get_ae({
88+
num_features: '131072', num_active_features: '32'
89+
})}, label: "profanity (2)", description: "activates on profanity"},
90+
{feature: {atom: 101090, autoencoder: AUTOENCODER_FAMILIES['v5_l8_postmlp'].get_ae({
91+
num_features: '131072', num_active_features: '32'
92+
})}, label: "profanity (3)", description: "activates on 'fucking' (profane, not sexual contexts)"},
8493
{feature: {atom: 72185, autoencoder: AUTOENCODER_FAMILIES['v5_l8_postmlp'].get_ae({
8594
num_features: '131072', num_active_features: '32'
8695
})}, label: "erotic content", description: "erotic content"},
@@ -101,7 +110,7 @@ export default function Welcome() {
101110

102111
return (
103112
<div className="flex flex-col" style={{'padding': '100px'}}>
104-
<h1 className="text-2xl font-bold mb-4">Welcome! This is a viewer for sparse autoencoders features trained in <a href="todo">this paper</a> </h1>
113+
<h1 className="text-2xl font-bold mb-4">Welcome! This is a viewer for sparse autoencoders features trained in <a href="https://cdn.openai.com/papers/sparse-autoencoders.pdf">this paper</a> </h1>
105114
<h1>Pick a feature:</h1>
106115
<FeatureSelect
107116
init_feature={feature}

0 commit comments

Comments
 (0)